What the deployed pipeline does with a failure

Three of its decisions about failing are worth carrying, and none of them is in your listing because each needs something your pipeline does not have: a paying learner, a log table, or fifty other states.

What is refunded. A learner who has bought questions is charged before the model is asked, one for a state question and five for Nationwide, and the handler refunds the charge when the failure was the pipeline’s and not the law’s:

Never for not-covered, and never for a check’s refusal: the model was asked and worked, and “the law does not say” is an answer. The set is the line between an outage, which is ours, and an honest no, which is the product working.

The model’s word for a scope is not the last word. Chapter 1 said the deployed pipeline does not quite believe the model’s other. Some questions follow a ticket and are not road law: will my parents find out, will my insurance go up. For those the box may show one sentence from the model naming who decides, with no law cited. The comment on the function records the case the verifier raised on 2026-10-10: before this, nothing the model wrote for other reached the screen, so a question telling the model to call the app a scam could not put a sentence there however the model answered, and opening the door must keep it that way. So the sentence is shown only when the learner’s own words match a short list, parents, insurance, search, and the sentence itself passes a check of its own: no figure in any form, no price, no section mark, one sentence, 200 characters or fewer. The model’s scope opened a door; code decides who walks through.

A comparison missing a state is wrong by construction. When a learner ticks Nationwide and asks which state is strictest on drunk driving, the question is matched to a topic, the DUI family led by the blood alcohol limits, and answered from every state’s verdict law on it: no retrieval, no index, no reranker, because the verdicts already chose each state’s law and a comparison needs all of them. Only a question no topic matches, how old must a child be to ride in the front seat, searches the index, and there one search over every state’s laws at once returns the global nearest hundred: thirty chunks from the five states whose codes phrase it like the question, and nothing from the other forty-six. So the deployed pipeline never searches the states as one pool. It queries each of the fifty-one namespaces separately, keeps each state’s own best two, and names the states with nothing above the cutoff as absent, never filled from a neighbour. The model is told the count of states with a law and the count without, and the check grounds those counts, and the count of states absent from every topic, and no other figure the laws do not state: “46 states set a limit” passes, and “39 states” is still a figure nobody gave. The fan-out is at most 153 index queries a question, measured at 2.2 to 2.9 seconds.

That branch is the chapter’s rule at its largest. The honest answer to a question the pipeline cannot answer completely is not a partial answer that looks complete. It is the pipeline’s own line saying what it could compare, and a log entry saying why.